A 600-case evaluation on the MuSiQue dataset using frozen Qwen2.5-7B-Instruct and MiniLM models tests whether bridge-aware context selection can outperform full context and other baselines.

  • Semantic bridge promotion achieved an F1 of 0.3482, surpassing BM25 (0.3113) and semantic core with random promotion (0.3310).
  • The approach used 16.6% fewer input tokens than full context while reducing prefill time by 15.3% and latency by 9.6%.
  • Paired bootstrap comparisons show statistically significant F1 gains over BM25 (+0.0368) and random controls (+0.0172).
  • The difference between bridge promotion and full context was not conclusive, with an F1 delta of +0.0101.

The results suggest that bridge-aware promotion preserves dependency-bearing evidence more reliably than generic shortening or lexical selection.